This removes the RTT-sampling function tfrc_tx_hist_rtt(), since
1. it suffered from complex passing of return values (the return value both
indicated successful lookup while the value doubled as RTT sample);
2. when for some odd reason the sample value equalled 0, this triggered a bug
warning about "bogus Ack", due to the ambiguity of the return value;
3. on a passive host which has not sent anything the TX history is empty and
thus will lead to unwanted "bogus Ack" warnings such as
ccid3_hc_tx_packet_recv: server(e7b7d518): DATAACK with bogus ACK-28197148
ccid3_hc_tx_packet_recv: server(e7b7d518): DATAACK with bogus ACK-26641606.
The fix is to replace the implicit encoding by performing the steps manually.
Furthermore, the "bogus Ack" warning has been removed, since it can actually be
triggered due to several reasons (network reordering, old packet, (3) above),
hence it is not very useful.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This fixes a subtle bug in the calculation of the inter-packet gap and shows
that t_delta, as it is currently used, is not needed.
The algorithm from RFC 5348, 8.3 below continually computes a send time t_nom,
which is initialised with the current time t_now; t_gran = 1E6 / HZ specifies
the scheduling granularity, s the packet size, and X the sending rate:
t_distance = t_nom - t_now; // in microseconds
t_delta = min(t_ipi, t_gran) / 2; // `delta' parameter in microseconds
if (t_distance >= t_delta) {
reschedule after (t_distance / 1000) milliseconds;
} else {
t_ipi = s / X; // inter-packet interval in usec
t_nom += t_ipi; // compute the next send time
send packet now;
}
Problem:
--------
Rescheduling requires a conversion into milliseconds (sk_reset_timer()). The
highest jiffy resolution with HZ=1000 is 1 millisecond, so using a higher
granularity does not make much sense here.
As a consequence, values of t_distance < 1000 are truncated to 0. This issue
has so far been resolved by using instead
if (t_distance >= t_delta + 1000)
reschedule after (t_distance / 1000) milliseconds;
This is unnecessarily large, a lower bound is t_delta' = max(t_delta, 1000).
And it implies a further simplification:
a) when HZ >= 500, then t_delta <= t_gran/2 = 10^6/(2*HZ) <= 1000, so that
t_delta' = MAX(1000, t_delta) = 1000 (constant value);
b) when HZ < 500, then t_delta = 1/2*MIN(rtt, t_ipi, t_gran) <= t_gran/2,
so that 1000 <= t_delta' <= t_gran/2.
The maximum error of using a constant t_delta in (b) is less than half a jiffy.
Fix:
----
The patch replaces t_delta with a constant, whose value depends on CONFIG_HZ,
changing the above algorithm to:
if (t_distance >= t_delta')
reschedule after (t_distance / 1000) milliseconds;
where t_delta' = 10^6/(2*HZ) if HZ < 500, and t_delta' = 1000 otherwise.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This makes RTAX_RTO_MIN also available to CCID-3, replacing the compile-time
RTO lower bound with a per-route tunable value.
The original Kconfig option solved the problem that a very low RTT (in the
order of HZ) can trigger too frequent and unnecessary reductions of the
sending rate.
This tunable does not affect the initial RTO value of 2 seconds specified in
RFC 5348, section 4.2 and Appendix B. But like the hardcoded Kconfig value,
it allows to adapt to network conditions.
The same effect as the original Kconfig option of 100ms is now achieved by
> ip route replace to unicast 192.168.0.0/24 rto_min 100j dev eth0
(assuming HZ=1000).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using a fixed RTO_MIN of 0.2 seconds was found to cause problems for CCID-2
over 802.11g: at least once per session there was a spurious timeout. It
helped to then increase the the value of RTO_MIN over this link.
Since the problem is the same as in TCP, this patch makes the solution from
commit "05bb1fad1cde025a864a90cfeb98dcbefe78a44a"
"[TCP]: Allow minimum RTO to be configurable via routing metrics."
available to DCCP.
This avoids reinventing the wheel, so that e.g. the following works in the
expected way now also for CCID-2:
> ip route change 10.0.0.2 rto_min 800 dev ath0
Luckily this useful rto_min function was recently moved to net/tcp.h,
which simplifies sharing code originating from TCP.
Documentation also updated (plus minor whitespace fixes).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch consolidates initial-window code common to TCP and CCID-2:
* TCP uses RFC 3390 in a packet-oriented manner (tcp_input.c) and
* CCID-2 uses RFC 3390 in packet-oriented manner (RFC 4341).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes the wrappers around the sk timer functions, since not much is
gained from using them: the BUG_ON in start_rto_timer will never trigger
since that function is called only if:
* the RTO timer expires (rto_expire, and then timer_pending() is false);
* in tx_packet_sent only if !timer_pending() (BUG_ON is redundant here);
* previously in new_ack, after stopping the timer (timer_pending() false).
Removing the wrappers also clears the way for eventually replacing the
RTO timer with the icsk-retransmission-timer, as it is already part of the
DCCP socket.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since CCID-2 is de facto a mini implementation of TCP, it makes sense to share
as much code as possible.
Hence this patch aligns CCID-2 timestamping with TCP timestamping.
This also halves the space consumption (on 64-bit systems).
The necessary include file <net/tcp.h> is already included by way of
net/dccp.h. Redundant includes have been removed.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current CCID-2 RTT estimator code is in parts broken and lags behind the
suggestions in RFC2988 of using scaled variants for SRTT/RTTVAR.
That code is replaced by the present patch, which reuses the Linux TCP RTT
estimator code.
Further details:
----------------
1. The minimum RTO of previously one second has been replaced with TCP's, since
RFC4341, sec. 5 says that the minimum of 1 sec. (suggested in RFC2988, 2.4)
is not necessary. Instead, the TCP_RTO_MIN is used, which agrees with DCCP's
concept of a default RTT (RFC 4340, 3.4).
2. The maximum RTO has been set to DCCP_RTO_MAX (64 sec), which agrees with
RFC2988, (2.5).
3. De-inlined the function ccid2_new_ack().
4. Added a FIXME: the RTT is sampled several times per Ack Vector, which will
give the wrong estimate. It should be replaced with one sample per Ack.
However, at the moment this can not be resolved easily, since
- it depends on TX history code (which also needs some work),
- the cleanest solution is not to use the `sent' time at all (saves 4 bytes
per entry) and use DCCP timestamps / elapsed time to estimated the RTT,
which however is non-trivial to get right (but needs to be done).
Reasons for reusing the Linux TCP estimator algorithm:
------------------------------------------------------
Some time was spent to find a better alternative, using basic RFC2988 as a first
step. Further analysis and experimentation showed that the Linux TCP RTO
estimator is superior to a basic RFC2988 implementation. A summary is on
http://www.erg.abdn.ac.uk/users/gerrit/dccp/notes/ccid2/rto_estimator/
In addition, this estimator fared well in a recent empirical evaluation:
Rewaskar, Sushant, Jasleen Kaur and F. Donelson Smith.
A Performance Study of Loss Detection/Recovery in Real-world TCP
Implementations. Proceedings of 15th IEEE International
Conference on Network Protocols (ICNP-07), 2007.
Thus there is significant benefit in reusing the existing TCP code.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes the dec_pipe function and improves the way the RTO timer is rearmed
when a new acknowledgment comes in.
Details and justification for removal:
--------------------------------------
1) The BUG_ON in dec_pipe is never triggered: pipe is only decremented for TX
history entries between tail and head, for which it had previously been
incremented in tx_packet_sent; and it is not decremented twice for the same
entry, since it is
- either decremented when a corresponding Ack Vector cell in state 0 or 1
was received (and then ccid2s_acked==1),
- or it is decremented when ccid2s_acked==0, as part of the loss detection
in tx_packet_recv (and hence it can not have been decremented earlier).
2) Restarting the RTO timer happens for every single entry in each Ack Vector
parsed by tx_packet_recv (according to RFC 4340, 11.4 this can happen up to
16192 times per Ack Vector).
3) The RTO timer should not be restarted when all outstanding data has been
acknowledged. This is currently done similar to (2), in dec_pipe, when
pipe has reached 0.
The patch onsolidates the code which rearms the RTO timer, combining the
segments from new_ack and dec_pipe. As a result, the code becomes clearer
(compare with tcp_rearm_rto()).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes the ccid2_hc_tx_check_sanity function: it is redundant.
Details:
The tx_check_sanity function performs three tests:
1) it checks that the circular TX list is sorted
- in ascending order of sequence number (ccid2s_seq)
- and time (ccid2s_sent),
- in the direction from `tail' (hctx_seqt) to `head' (hctx_seqh);
2) it ensures that the entire list has the length seqbufc * CCID2_SEQBUF_LEN;
3) it ensures that pipe equals the number of packets that were not
marked `acked' (ccid2s_acked) between `tail' and `head'.
The following argues that each of these tests is redundant, this can be verified
by going through the code.
(1) is not necessary, since both time and GSS increase from one packet to the
next, so that subsequent insertions in tx_packet_sent (which advance the `head'
pointer) will be in ascending order of time and sequence number.
In (2), the length of the list is always equal to seqbufc times CCID2_SEQBUF_LEN
(set to 1024) unless allocation caused an earlier failure, because:
* at initialisation (tx_init), there is one chunk of size 1024 and seqbufc=1;
* subsequent calls to tx_alloc_seq take place whenever head->next == tail in
tx_packet_sent; then a new chunk of size 1024 is inserted between head and
tail, and seqbufc is incremented by one.
To show that (3) is redundant requires looking at two cases.
The `pipe' variable of the TX socket is incremented only in tx_packet_sent, and
decremented in tx_packet_recv. When head == tail (TX history empty) then pipe
should be 0, which is the case directly after initialisation and after a
retransmission timeout has occurred (ccid2_hc_tx_rto_expire).
The first case involves parsing Ack Vectors for packets recorded in the live
portion of the buffer, between tail and head. For each packet marked by the
receiver as received (state 0) or ECN-marked (state 1), pipe is decremented by
one, so for all such packets the BUG_ON in tx_check_sanity will not trigger.
The second case is the loss detection in the second half of tx_packet_recv,
below the comment "Check for NUMDUPACK".
The first while-loop here ensures that the sequence number of `seqp' is either
above or equal to `high_ack', or otherwise equal to the highest sequence number
sent so far (of the entry head->prev, as head points to the next unsent entry).
The next while-loop ("while (1)") counts the number of acked packets starting
from that position of seqp, going backwards in the direction from head->prev to
tail. If NUMDUPACK=3 such packets were counted within this loop, `seqp' points
to the last acknowledged packet of these, and the "if (done == NUMDUPACK)" block
is entered next.
The while-loop contained within that block in turn traverses the list backwards,
from head to tail; the position of `seqp' is saved in the variable `last_acked'.
For each packet not marked as `acked', a congestion event is triggered within
the loop, and pipe is decremented. The loop terminates when `seqp' has reached
`tail', whereupon tail is set to the position previously stored in `last_acked'.
Thus, between `last_acked' and the previous position of `tail',
- pipe has been decremented earlier if the packet was marked as state 0 or 1;
- pipe was decremented if the packet was not marked as acked.
That is, pipe has been decremented by the number of packets between `last_acked'
and the previous position of `tail'. As a consequence, pipe now again reflects
the number of packets which have not (yet) been acked between the new position
of tail (at `last_acked') and head->prev, or 0 if head==tail. The result is that
the BUG_ON condition in check_sanity will also not be triggered, hence the test
(3) is also redundant.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The CCIDs are activated as last of the features, at the end of the handshake,
were the LISTEN state of the master socket is inherited into the server
state of the child socket. Thus, the only states visible to CCIDs now are
OPEN/PARTOPEN, and the closing states.
This allows to remove tests which were previously necessary to protect
against referencing a socket in the listening state (in CCID-3), but which
now have become redundant.
As a further byproduct of enabling the CCIDs only after the connection has been
fully established, several typecast-initialisations of ccid3_hc_{rx,tx}_sock
can now be eliminated:
* the CCID is loaded, so it is not necessary to test if it is NULL,
* if it is possible to load a CCID and leave the private area NULL, then this
is a bug, which should crash loudly - and earlier,
* the test for state==OPEN || state==PARTOPEN now reduces only to the closing
phase (e.g. when the node has received an unexpected Reset).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch collects cosmetics-only changes to separate these from
code changes:
* update with regard to CodingStyle and whitespace changes,
* documentation:
- adding/revising comments,
- remove CCID-3 RX socket documentation which is either
duplicate or refers to fields that no longer exist,
* expand embedded tfrc_tx_info struct inline for consistency,
removing indirections via #define.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes an unused 'sk' argument from several option-inserting functions.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
This continues the previous patch, by applying the same change to CCID-3.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This removes a redundancy in the CCID half-connection (hc) naming scheme:
* instead of 'hctx->tx_...', write 'hc->tx_...';
* instead of 'hcrx->rx_...', write 'hc->rx_...';
which works because the 'type' of the half-connection is encoded in the
'rx_' / 'tx_' prefixes.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This implements the new naming scheme also for CCID-3.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch starts a less problematic naming convention for CCID structs.
The old naming convention used 'hc{tx,rx}->ccid?hc{tx,rx}->...' as
recurring prefixes, which made the code
* hard to write (not easy to fit into 80 characters);
* hard to read (most of the space is occupied by prefixes).
The new naming scheme:
* struct entries for the TX socket are prefixed by 'tx_';
* and those for the RX socket are prefixed by 'rx_'.
The identifiers then remain distinguishable when grep-ing through the tree:
(a) RX/TX sockets are distinguished by the naming scheme,
(b) individual CCIDs are distinguished by filename (ccid{2,3,4}.{c,h}).
This first patch implements the scheme for CCID-2.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
No code change, cosmetical changes only:
* whitespace cleanup via scripts/cleanfile,
* remove self-references to filename at top of files,
* fix coding style (extraneous brackets),
* fix documentation style (kernel-doc-nano-HOWTO).
Thanks are due to Ivo Augusto Calado who raised these issues by
submitting good-quality patches.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
String literals are constant, and usually, we can also tag the array
of pointers const too, moving it to the .rodata section.
Signed-off-by: Jan Engelhardt <jengelh@medozas.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Thanks to Wei and Arnaldo for pointing out the correct
new reference for CCID-3.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Removed the __exit annotation of tfrc_lib_exit(), in order to suppress the following section mismatch messages:
WARNING: net/dccp/dccp.o(.text+0xd9): Section mismatch in reference from the function ccid_cleanup_builtins() to the function .exit.text:tfrc_lib_exit()
The function ccid_cleanup_builtins() references a function in an exit section.
Often the function tfrc_lib_exit() has valid usage outside the exit section
and the fix is to remove the __exit annotation of tfrc_lib_exit.
WARNING: net/dccp/dccp.o(.init.text+0x48): Section mismatch in reference from the function ccid_initialize_builtins() to the function .exit.text:tfrc_lib_exit()
The function __init ccid_initialize_builtins() references
a function __exit tfrc_lib_exit().
This is often seen when error handling in the init function
uses functionality in the exit path.
The fix is often to remove the __exit annotation of
tfrc_lib_exit() so it may be used outside an exit section.
Signed-off-by: Leonardo Potenza <lpotenza@inwind.it>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch integrates the TFRC library, which is a dependency of CCID-3 (and
CCID-4), with the new use of CCIDs in the DCCP module.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on Arnaldo's earlier patch, this patch integrates the standardised
CCID congestion control plugins (CCID-2 and CCID-3) of DCCP with dccp.ko:
* enables a faster connection path by eliminating the need to always go
through the CCID registration lock;
* updates the implementation to use only a single array whose size equals
the number of configured CCIDs instead of the maximum (256);
* since the CCIDs are now fixed array elements, synchronization is no
longer needed, simplifying use and implementation.
CCID-2 is suggested as minimum for a basic DCCP implementation (RFC 4340, 10);
CCID-3 is a standards-track CCID supported by RFC 4342 and RFC 5348.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Two registration routines, for SP and NN features, are provided by this patch,
replacing a previous routine which was used for both feature types.
These are internal-only routines and therefore start with `__feat_register'.
It further exports the known limits of Sequence Window and Ack Ratio as symbolic
constants.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Ian McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: David S. Miller <davem@davemloft.net>
This implements [RFC 3448, 4.5], which performs congestion avoidance behaviour
by reducing the transmit rate as the queueing delay (measured in terms of
long-term RTT) increases.
Oscillation can be turned on/off via a module option (do_osc_prev) and via sysfs
(using mode 0644), the default is off.
Overflow analysis:
------------------
* oscillation prevention is done after update_x(), so that t_ipi <= 64000;
* hence the multiplication "t_ipi * sqrt(R_sample)" needs 64 bits;
* done using u64 for sqrt_sample and explicit typecast of t_ipi;
* the divisor, R_sqmean, is non-zero because oscillation prevention is first
called when receiving the second feedback packet, and tfrc_scaled_rtt() > 0.
A detailed discussion of the algorithm (with plots) is on
http://www.erg.abdn.ac.uk/users/gerrit/dccp/notes/ccid3/sender_notes/oscillation_prevention/
The algorithm has negative side effects:
* when allowing to decrease t_ipi (leads to a large RTT) and
* when using it during slow-start;
both uses are therefore disabled.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This patch simplifies the computation of t_ipi, avoiding expensive computations
to enforce the minimum sending rate.
Both RFC 3448 and rfc3448bis (revision #06), as well as RFC 4342 sec 5., require
at various stages that at least one packet must be sent per t_mbi = 64 seconds.
This requires frequent divisions of the type X_min = s/t_mbi, which are later
converted back into an inter-packet-interval t_ipi_max = s/X_min = t_mbi.
The patch removes the expensive indirection; in the unlikely case of having
a sending rate less than one packet per 64 seconds, it also re-adjusts X.
The following cases document conformance with RFC 3448 / rfc3448bis-06:
1) Time until receiving the first feedback packet:
* if the sender has no initial RTT sample then X = s/1 Bps > s/t_mbi;
* if the sender has an initial RTT sample or when the first feedback
packet is received, X = W_init/R > s/t_mbi.
2) Slow-start (p == 0 and feedback packets come in):
* RFC 3448 (current code) enforces a minimum of s/R > s/t_mbi;
* rfc3448bis (future code) enforces an even higher minimum of W_init/R.
3) Congestion avoidance with no absence of feedback (p > 0):
* when X_calc or X_recv/2 are too low, the minimum of X_min = s/t_mbi
is enforced in update_x() when calling update_send_interval();
* update_send_interval() is, as before, only called when X changes
(i.e. either when increasing or decreasing, not when in equilibrium).
4) Reduction of X without prior feedback or during slow-start (p==0):
* both RFC 3448 and rfc3448bis here halve X directly;
* the associated constraint X >= s/t_mbi is nforced here by send_interval().
5) Reduction of X when p > 0:
* X is modified indirectly via X_recv (RFC 3448) or X_recv_set (rfc3448bis);
* in both cases, control goes back to section 4.3 (in both documents);
* since p > 0, both documents use X = max(min(...), s/t_mbi), which is
enforced in this patch by calling send_interval() from update_x().
I think that this analysis is exhaustive. Should I have forgotten a case,
the worst-case consideration arises when X sinks below s/t_mbi, and is then
increased back up to this minimum value. Even under this assumption, the
behaviour is correct, since all lower limits of X in RFC 3448 / rfc3448bis
are either equal to or greater than s/t_mbi.
Note on the condition X >= s/t_mbi <==> t_ipi = s/X <= t_mbi: since X is
scaled by 64, and all time units are in microseconds, the coded condition is:
t_ipi = s * 64 * 10^6 usec / X <= 64 * 10^6 usec
This simplifies to s / X <= 1 second <==> X * 1 second >= s > 0.
(A zero `s' is not allowed by the CCID-3 code).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
rfc3448bis allows three different ways of tracking the packet size `s':
1. using the MSS/MPS (at initialisation, 4.2, and in 4.1 (1));
2. using the average of `s' (in 4.1);
3. using the maximum of `s' (in 4.2).
Instead of hard-coding a single interpretation of rfc3448bis, this implements
a choice of all three alternatives and suggests the first as default, since it
is the option which is most consistent with other parts of the specification.
The patch further deprecates the update of t_ipi whenever `s' changes. The
gains of doing this are only small since a change of s takes effect at the
next instant X is updated:
* when the next feedback comes in (within one RTT or less);
* when the nofeedback timer expires (within at most 4 RTTs).
Further, there are complications caused by updating t_ipi whenever s changes:
* if t_ipi had previously been updated to effect oscillation prevention (4.5),
then it is impossible to make the same adjustment to t_ipi again, thus
counter-acting the algorithm;
* s may be updated any time and a modification of t_ipi depends on the current
state (e.g. no oscillation prevention is done in the absence of feedback);
* in rev-06 of rfc3448bis, there are more possible cases, depending on whether
the sender is in slow-start (t_ipi <= R/W_init), or in congestion-avoidance,
limited by X_recv or the throughput equation (t_ipi <= t_mbi).
Thus there are side effects of always updating t_ipi as s changes. These may not
be desirable. The only case I can think of where such an update makes sense is
to recompute X_calc when p > 0 and when s changes (not done by this patch).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
The per-CCID menu has several dependencies on EXPERIMENTAL. These are redundant,
since net/dccp/ccids/Kconfig is sourced by net/dccp/Kconfig and since the
latter menu in turn asserts a dependency on EXPERIMENTAL.
The patch removes the redundant dependencies as well as the repeated reference
within the sub-menu.
Further changes:
----------------
Two single dependencies on CCID-3 are replaced with a single enclosing `if'.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
The patch updates CCID-3 with regard to the latest rfc3448bis-06:
* in the first revisions of the draft, MSS was used for the RFC 3390 window;
* then (from revision #1 to revision #2), it used the packet size `s';
* now, in this revision (and apparently final), the value is back to MSS.
This change has an implication for the case when no RTT sample is available,
at the time of sending the first packet:
* with RTT sample, 2*MSS/RTT <= initial_rate <= 4*MSS/RTT;
* without RTT sample, the initial rate is one packet (s bytes) per second
(sec. 4.2), but using s instead of MSS here creates an imbalance, since
this would further reduce the initial sending rate.
Hence the patch uses MSS (called MPS in RFC 4340) in all places.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This patch is a requirement for enabling ECN support later on. With that change
in mind, the following preparations are done:
* renamed handle_loss() into congestion_event() since it returns true when a
congestion event happens (it will eventually also take care of ECN packets);
* lets tfrc_rx_congestion_event() always update the RX history records, since
this routine needs to be called for each non-duplicate packet anyway;
* made all involved boolean-type functions to have return type `bool';
Updating the RX history records is now only necessary for the packets received
up to sending the first feedback. The receiver code becomes again simpler.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This updates the computation of X_recv with regard to Errata 610/611 for
RFC 4342 and draft rfc3448bis-06, ensuring that at least an interval of 1
RTT is used to compute X_recv. The change is wrapped into a new function
ccid3_hc_rx_x_recv().
Further changes:
----------------
* feedback is not sent when no data packets arrived (bytes_recv == 0), as per
rfc3448bis-06, 6.2;
* take the timestamp for the feedback /after/ dccp_send_ack() returns, to avoid
taking the transmission time into account (in case layer-2 is busy);
* clearer handling of failure in ccid3_first_li().
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This improves the receiver RTT sampling algorithm so that it tries harder to get
as many RTT samples as possible.
The algorithm is based the concepts presented in RFC 4340, 8.1, using timestamps
and the CCVal window counter. There exist 4 cases for the CCVal difference:
* == 0: less than RTT/4 passed since last packet -- unusable;
* > 4: (much) more than 1 RTT has passed since last packet -- also unusable;
* == 4: perfect sample (exactly one RTT has passed since last packet);
* 1..3: sub-optimal sample (between RTT/4 and 3*RTT/4 has passed).
In the last case the algorithm tried to optimise by storing away the candidate
and then re-trying next time. The problem is that
* a large number of samples is needed to smooth out the inaccuracies of the
algorithm;
* the sender may not be sending enough packets to warrant a "next time";
* hence it is better to use suboptimal samples whenever possible.
The algorithm now stores away the current sample only if the difference is 0.
Applicability and background
----------------------------
A realistic example is MP3 streaming where packets are sent at a rate of less
than one packet per RTT, which means that suitable samples are absent for a
very long time.
The effectiveness of using suboptimal samples (with a delta between 1 and 4) was
confirmed by instrumenting the algorithm with counters. The results of two 20
second test runs were:
* With the old algorithm and a total of 38442 function calls, only 394 of these
calls resulted in usable RTT samples (about 1%), and 378 out of these were
"perfect" samples and 28013 (unused) samples had a delta of 1..3.
* With the new algorithm and a total of 37057 function calls, 1702 usable RTT
samples were retrieved (about 4.6%), 5 out of these were "perfect" samples.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This updates the CCID-3 receiver in part with regard to errata 610 and 611
(http://www.rfc-editor.org/errata_list.php), which change RFC 4342 to use the
Receive Rate as specified in rfc3448bis, requiring to constantly sample the
RTT (or use a sender RTT).
Doing this requires reusing the RX history structure after dealing with a loss.
The patch does not resolve how to compute X_recv if the interval is less
than 1 RTT. A FIXME has been added (and is resolved in subsequent patch).
Furthermore, since this is all TFRC-based functionality, the RTT estimation
is now also performed by the dccp_tfrc_lib module. This further simplifies
the CCID-3 code.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
The only state information that the CCID-3 receiver keeps is whether initial
feedback has been sent or not. Further, this overlaps with use of feedback:
* state == TFRC_RSTATE_NO_DATA as long as no feedback has been sent;
* state == TFRC_RSTATE_DATA as soon as the first feedback has been sent.
This patch reduces the duplication, by memorising the type of the last feedback.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This migrates more TFRC-related code into the dccp_tfrc_lib:
* sampling of the packet size `s' (which is only needed until the first
loss interval is computed (ccid3_first_li));
* updating the byte-counter `bytes_recvd' in between sending feedbacks.
The result is a better separation of CCID-3 specific and TFRC specific
code, which aids future integration with ECN and e.g. CCID-4.
Further changes:
----------------
* replaced magic number of 536 with equivalent constant TCP_MIN_RCVMSS;
(this constant is also used when no estimate for `s' is available).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This changes the return type of tfrc_lh_update_i_mean() to void, since that
function returns always `false'. This is due to
len = dccp_delta_seqno(cur->li_seqno, DCCP_SKB_CB(skb)->dccpd_seq) + 1;
if (len - (s64)cur->li_length <= 0) /* duplicate or reordered */
return 0;
which means that update_i_mean can only increase the length of the open loss
interval I_0, and hence the value of I_tot0 (RFC 3448, 5.4). Consequently the
test `i_mean < old_i_mean' at the end of the function always evaluates to false.
There is no known way by which a loss interval can suddenly become shorter,
therefore the return type of the function is changed to void. (That is, under
the given circumstances step (3) in RFC 3448, 6.1 will not occur.)
Further changes:
----------------
* the function is now called from tfrc_rx_handle_loss, which is equivalent
to the previous way of calling from rx_packet_recv (it was called whenever
there was no new or pending loss, now it is also updated when there is
a pending loss - this increases the accuracy a bit);
* added a FIXME to possibly consider NDP counting as per RFC 4342 (this is
not implemented yet).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This enables the TFRC code to begin loss detection (as soon as the module
is loaded), using the latest updates from rfc3448bis-06, 6.3.1:
* when the first data packet(s) are lost or marked, set
* X_target = s/(2*R) => f(p) = s/(R * X_target) = 2,
* corresponding to a loss rate of ~ 20.64%.
The handle_loss() function is now called right at the begin of rx_packet_recv()
and thus no longer protected against duplicates: hence a call to rx_duplicate()
has been added. Such a call makes sense now, as the previous patch initialises
the first entry with a sequence number of GSR.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This patch
1) separates history allocation and initialisation, to facilitate early
loss detection (implemented by a subsequent patch);
2) removes duplication by using the existing tfrc_rx_hist_purge() if the
allocation fails. This is now possible, since the initialisation routine
3) zeroes out the entire history before using it.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
In the congestion-avoidance phase a decay of p towards 0 is natural once fewer
losses are encountered. Hence the warning message "p is below resolution" is
not necessary, and thus turned into a debug message by this patch.
The TFRC_SMALLEST_P is needed since in theory p never actually reaches 0. When
no further losses are encountered, the loss interval I_0 grows in length,
causing p to decrease towards 0, causing X_calc = s/(RTT * f(p)) to increase.
With the given minimum-resolution this congestion avoidance phase stops at some
fixed value, an approximation formula has been added to the documentation.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Since CCIDs are only used during the established phase of a connection,
they have very little internal state; this specifically reduces to:
* "no packet sent" if and only if s == 0, for the TX packet size s;
* when the first packet has been sent (i.e. `s' > 0), the question is whether
or not feedback has been received:
- if a feedback packet is received, "feedback = yes" is set,
- if the nofeedback timer expires, "feedback = no" is set.
Thus the CCID only needs to remember state about whether or not feedback
has been received. This is now implemented using a boolean flag, which is
toggled when a feedback packet arrives or the nofeedback timer expires.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
The DCCP base time resolution is 10 microseconds (RFC 4340, 13.1 ... 13.3).
Using a timer with a lower resolution was found to trigger the following
bug warnings/problems on high-speed networks (e.g. local loopback):
* RTT samples are rounded down to 0 if below resolution;
* in some cases, negative RTT samples were observed;
* the CCID-3 feedback timer complains that the feedback interval is 0,
since the feedback interval is in the order of 1 RTT or less and RTT
measurement rounded this down to 0;
On an Intel computer this will for instance happen when using a
boot-time parameter of "clocksource=jiffies".
The following system log messages were observed:
11:24:00 kernel: BUG: delta (0) <= 0 at ccid3_hc_rx_send_feedback()
11:26:12 kernel: BUG: delta (0) <= 0 at ccid3_hc_rx_send_feedback()
11:26:30 kernel: dccp_sample_rtt: unusable RTT sample 0, using min
11:26:30 last message repeated 5 times
This patch defines a global constant for the time resolution, adds this in
timer.c, and checks the available clock resolution at CCID-3 module load time.
When the resolution is worse than 10 microseconds, module loading exits with
a message "socket type not supported".
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This patch consolidates the code common to TCP and CCID-2:
* TCP uses RFC 3390 in a packet-oriented manner (tcp_input.c) and
* CCID-2 uses RFC 3390 in packet-oriented manner (RFC 4341).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This removes the wrappers around the sk timer functions as it makes the code
clearer and not much is gained from using wrappers: the BUG_ON in
start_rto_timer will never trigger since that function was called only when
* the RTO timer expired (rto_expire, and then timer_pending() is false);
* in tx_packet_sent only if !timer_pending() (BUG_ON is redundant here);
* previously in new_ack, after stopping the timer (timer_pending() false).
One further motive behind this patch is to replace the RTO timer with the
icsk retransmission timer, as it is already part of the DCCP socket.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
The current CCID-2 RTT estimator code is in parts broken and lags behind the
suggestions in RFC2988 of using scaled variants for SRTT/RTTVAR.
That code is replaced by the present patch, which reuses the Linux TCP RTT
estimator code - reasons for this code duplication are given below.
Further details:
----------------
1. The minimum RTO of previously one second has been replaced with TCP's, since
RFC4341, sec. 5 says that the minimum of 1 sec. (suggested in RFC2988, 2.4)
is not necessary. Instead, the TCP_RTO_MIN is used, which agrees with DCCP's
concept of a default RTT (RFC 4340, 3.4).
2. The maximum RTO has been set to DCCP_RTO_MAX (64 sec), which agrees with
RFC2988, (2.5).
3. De-inlined the function ccid2_new_ack().
4. Added a FIXME: the RTT is sampled several times per Ack Vector, which will
give the wrong estimate. It should be replaced with one sample per Ack.
However, at the moment this can not be resolved easily, since
- it depends on TX history code (which also needs some work),
- the cleanest solution is not to use the `sent' time at all (saves 4 bytes
per entry) and use DCCP timestamps / elapsed time to estimated the RTT,
which however is non-trivial to get right (but needs to be done).
Reasons for reusing the Linux TCP estimator algorithm:
------------------------------------------------------
Some time was spent to find a better alternative, using basic RFC2988 as a first
step. Further analysis and experimentation showed that the Linux TCP RTO
estimator is superior to a basic RFC2988 implementation. A summary is on
http://www.erg.abdn.ac.uk/users/gerrit/dccp/notes/ccid2/rto_estimator/
In addition, this estimator fared well in a recent empirical evaluation:
Rewaskar, Sushant, Jasleen Kaur and F. Donelson Smith.
A Performance Study of Loss Detection/Recovery in Real-world TCP
Implementations. Proceedings of 15th IEEE International
Conference on Network Protocols (ICNP-07). 2007.
Thus there is significant benefit in reusing the existing TCP code.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This removes the dec_pipe function and improves the way the RTO timer is rearmed
when a new acknowledgment comes in.
Details and justification for removal:
--------------------------------------
1) The BUG_ON in dec_pipe is never triggered: pipe is only decremented for TX
history entries between tail and head, for which it had previously been
incremented in tx_packet_sent; and it is not decremented twice for the same
entry, since it is
- either decremented when a corresponding Ack Vector cell in state 0 or 1
was received (and then ccid2s_acked==1),
- or it is decremented when ccid2s_acked==0, as part of the loss detection
in tx_packet_recv (and hence it can not have been decremented earlier).
2) Restarting the RTO timer happens for every single entry in each Ack Vector
parsed by tx_packet_recv (according to RFC 4340, 11.4 this can happen up to
16192 times per Ack Vector).
3) The RTO timer should not be restarted when all outstanding data has been
acknowledged. This is currently done similar to (2), in dec_pipe, when
pipe has reached 0.
The patch onsolidates the code which rearms the RTO timer, combining the
segments from new_ack and dec_pipe. As a result, the code becomes clearer
(compare with tcp_rearm_rto()).
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This removes the ccid2_hc_tx_check_sanity function: it is redundant.
Details:
========
The tx_check_sanity function performs three tests:
1) it checks that the circular TX list is sorted
- in ascending order of sequence number (ccid2s_seq)
- and time (ccid2s_sent),
- in the direction from `tail' (hctx_seqt) to `head' (hctx_seqh);
2) it ensures that the entire list has the length seqbufc * CCID2_SEQBUF_LEN;
3) it ensures that pipe equals the number of packets that were not
marked `acked' (ccid2s_acked) between `tail' and `head'.
The following argues that each of these tests is redundant, this can be verified
by going through the code.
(1) is not necessary, since both time and GSS increase from one packet to the
next, so that subsequent insertions in tx_packet_sent (which advance the `head'
pointer) will be in ascending order of time and sequence number.
In (2), the length of the list is always equal to seqbufc times CCID2_SEQBUF_LEN
(set to 1024) unless allocation caused an earlier failure, because:
* at initialisation (tx_init), there is one chunk of size 1024 and seqbufc=1;
* subsequent calls to tx_alloc_seq take place whenever head->next == tail in
tx_packet_sent; then a new chunk of size 1024 is inserted between head and
tail, and seqbufc is incremented by one.
To show that (3) is redundant requires looking at two cases.
The `pipe' variable of the TX socket is incremented only in tx_packet_sent, and
decremented in tx_packet_recv. When head == tail (TX history empty) then pipe
should be 0, which is the case directly after initialisation and after a
retransmission timeout has occurred (ccid2_hc_tx_rto_expire).
The first case involves parsing Ack Vectors for packets recorded in the live
portion of the buffer, between tail and head. For each packet marked by the
receiver as received (state 0) or ECN-marked (state 1), pipe is decremented by
one, so for all such packets the BUG_ON in tx_check_sanity will not trigger.
The second case is the loss detection in the second half of tx_packet_recv,
below the comment "Check for NUMDUPACK".
The first while-loop here ensures that the sequence number of `seqp' is either
above or equal to `high_ack', or otherwise equal to the highest sequence number
sent so far (of the entry head->prev, as head points to the next unsent entry).
The next while-loop ("while (1)") counts the number of acked packets starting
from that position of seqp, going backwards in the direction from head->prev to
tail. If NUMDUPACK=3 such packets were counted within this loop, `seqp' points
to the last acknowledged packet of these, and the "if (done == NUMDUPACK)" block
is entered next.
The while-loop contained within that block in turn traverses the list backwards,
from head to tail; the position of `seqp' is saved in the variable `last_acked'.
For each packet not marked as `acked', a congestion event is triggered within
the loop, and pipe is decremented. The loop terminates when `seqp' has reached
`tail', whereupon tail is set to the position previously stored in `last_acked'.
Thus, between `last_acked' and the previous position of `tail',
- pipe has been decremented earlier if the packet was marked as state 0 or 1;
- pipe was decremented if the packet was not marked as acked.
That is, pipe has been decremented by the number of packets between `last_acked'
and the previous position of `tail'. As a consequence, pipe now again reflects
the number of packets which have not (yet) been acked between the new position
of tail (at `last_acked') and head->prev, or 0 if head==tail. The result is that
the BUG_ON condition in check_sanity will also not be triggered, hence the test
(3) is also redundant.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
